Embed Size (px)
Citation preview
1
Qualitative and quantitative aspects of phonetic variation in Dutch
eigenlijk
Mirjam Ernestusa,b & Rachel Smithc
aCentre for Language studies, Radboud University, Nijmegen, The Netherlands bMax Planck Institute for Psycholingusitics, Nijmegen, The Netherlands cGlasgow University Laboratory of Phonetics, University of Glasgow, UK
1. Introduction
In informal conversations in many languages, many word tokens are pronounced
with fewer segments or with segments that are articulated more weakly than in
careful speech (for an introduction to the phenomenon, see Ernestus and Warner
2011). For instance, the word particular may be pronounced like [phthιkhә] and
hilarious like [hlɛrɛs] (Johnson 2004). These short word pronunciation variants
are generally referred to as reduced forms, and we adopt this terminology here.
Reduced forms typically occur in weak prosodic positions, especially in
unaccented positions in the middle of sentences (e.g. Pluymaekers, Ernestus &
Baayen. 2005a). This paper contributes to our knowledge of the characteristics
of reduced forms by studying in detail one word type in Dutch (i.e. eigenlijk
'actually') that is known to show wide variation in its realization (e.g. Ernestus
2000: 141). The results shed light on the variation that a word may show and on
how speakers from the same sociolinguistic group may differ in how they reduce.
In addition, the results raise questions about the nature of reduced forms, the
mental lexicon and psycholinguistic models of speech production and
comprehension.
Nearly all previous research on reduced forms focused on the presence
versus absence of segments and on the duration of these segments or (parts of)
the words as measures of reduction. These studies have shown that many
different factors affect the probability that a given word appears in a reduced
form. These factors include speech rate (e.g. Raymond, Dautricourt, and Hume
2006; Kohler 1990), the word's phonological neighborhood density (Gahl, Yao,
and Johnson 2012), its prosodic position (e.g. Bell et al. 2003), its a-priori-
probability (e.g. Pluymaekers, Ernestus and Baayen 2005a; Gahl 2008), its
probability in context (e.g. Bell et al. 2003; Pluymaekers, Ernestus, and Baayen
2005b; Bell et al. 2009) and the presence versus absence of a following
hesitation (e.g., Bell et al. 2003). The influences of these factors suggest that
reduction may result from time pressure: when time pressure is high, for
instance because speech rate is high or because the word or the following word
was easy to plan and is ready to be articulated, reduction is more likely to occur
(e.g. Bell et al. 2009; Gahl et al. 2012).
In addition, several studies focusing on duration and on the presence
versus absence of segments suggest that degree of reduction is under the
speaker's direct control. For instance, speakers may choose not to reduce at
high speech rates (e.g. van Son and Pols 1990, 1992), and degree of reduction
2
correlates with speaker characteristics, including gender (e.g. Guy 1991; Phillips
1994), age (e.g. Guy 1991; Strik, van Doremalen, and Cucchiarini 2008) and
socio-economic status (e.g. Labov 2001). Furthermore, speakers of different
regiolects of a language may differ in degree of reduction for some words (e.g.
Keune et al. 2005). Reduction is therefore not a fully automatic process, but is
speaker-dependent and probably at least partly under the speaker’s control.
Only a few studies so far have investigated more detailed phonetic
characteristics of reduced forms. Such studies have shown that information
about a word’s identity is often preserved despite reduction or reorganization of
the acoustic features or articulatory gestures that would be found in a canonical
form. For example, reduced tokens of support may lack any evidence of a vowel
portion between /s/ and the closure of /p/, yet may maintain aspiration of /p/,
which is consistent with a singleton syllable-initial stop, rather than an /sp/
cluster. Thus the laryngeal specification of the stop preserves information that
prevents reduced support from becoming ambiguous with sport (Manuel 1991;
Manuel et al. 1992; see also Davidson 2006, and see Aalders & Ernestus, in
preparation, for evidence that this also holds in casual speech). Similarly, some
reduced forms of French c’était 'it was' can lack a voiced vowel between /s/ and
/t/, yet retain traces of the vowel in the form of a lowered spectral centre of
gravity in the latter part of the /s/ (Torreira and Ernestus 2011). In English the,
/ð/ can assimilate in manner of articulation to a preceding nasal or lateral (in
phrases like in the, all the), losing any evidence of frication; yet residues of /ð/
tend to be retained in the form of dentality (as cued by F2 at the nasal or lateral
boundaries) and duration (Manuel 1995).
In extreme cases of reduction it may be impossible to linearly segment
the speech signal, yet sufficient phonetic residue of a word’s form may remain as
to make it fully identifiable. Kohler (1999) described such residues as
“articulatory prosodies”, which “persist as non-linear, suprasegmental features
of syllables, reflecting, e.g., nasality or labiality that is no longer tied to specific
segmental units’’ and may be quite extended in time (p. 89). For example, the
German discourse marker eigentlich 'actually' is canonically produced as
[aɪgŋtliç], but can be reduced to [aɪŋi] or [aɪi], with palatality, nasality, and
duration serving to convey the word’s “phonetic essence” (Niebuhr and Kohler
2011). Perception tests indicate that listeners may be sensitive to such
articulatory prosodies, even in the absence of contextual clues (Niebuhr and
Kohler 2011), just as they are to other aspects of phonetic detail in reduced
speech (Manuel 1991, 1995). Thus reduction may involve significant departures
from a word’s canonical form, while preserving phonological contrast quite well
(Warner and Tucker 2011).
The degree of reduction may be affected by the function that a word
performs. Plug (2005) investigated reduction of the Dutch discourse marker
eigenlijk 'actually' as a function of its pragmatic function. Plug analysed 49
tokens of eigenlijk performing two of the word’s subfunctions, one being
correction or clarification of a statement or assumption in a speaker’s own
utterance (self-repair), and the other being correction or clarification of
3
something said or assumed by the interlocutor (other-repair). He observed that
tokens with the function of self-repair tended to be produced fast and to be
highly reduced in terms of their number of syllables and segments. Tokens
whose function was other-repair tended to occur at the edges of prosodic
phrases and to be produced slower and with more phonetic elaboration.
Like Plug (2005), the present study focusses on the Dutch word eigenlijk.
As is the case for nearly all words in every language, we have very little detailed
knowledge about the possible pronunciation variants of the word, and about how
frequently these variants occur and under which conditions. The present study
examines over 150 tokens of this word, produced by 18 speakers in informal
conversations, examining their detailed phonetic characteristics and when
particular clusters of characteristics are most likely to occur. Detailed data on
the pronunciation variation of this word will form a good testing ground for
common assumptions about reduction, including the assumption that reduced
forms only occur in prosodically weak positions, and the related assumption that
speakers mainly reduce to cope with time pressure.
Our main reason for studying eigenlijk is that it is known to show a wide
variation in pronunciation, ranging from trisyllablic /'ɛιxәlәk/ (see Figure 1 for an
example) to monosyllabic variants like /'ɛιxk/ and /'ɛιk/ (see e.g. Ernestus 2000;
Plug 2005; Pluymaekers, Ernestus and Baayen 2005b). The word shares this
variability with many other words also ending in the suffix -/lәk/ -lijk (e.g.
Pluymaekers, Ernestus and Baayen 2005b). The word occurs relatively
frequently in informal conversations (e.g. 1922 tokens per million word tokens in
the Spoken Dutch Corpus, Oostdijk 2000), which allows us to study both intra-
and interspeaker variation on the basis of tokens produced in a relatively short
period of time.
As mentioned above, the word eigenlijk is a discourse marker that in
general signals a contrast between what the speaker is saying and what (s)he or
the interlocutor just said or implied, or is assumed to believe (e.g. Plug 2005;
van Bergen et al. 2011); see, for instance, sentences (1, 2, 3) from our data set.
(1) een van de, of eigenlijk de oudste, acht geveild zou worden.
one of the, or actually the oldest, eight [a type of rowing boat]
would be auctioned.
(2) Nee, tenten hoef ik eigenlijk niet
No, I actually do not need tents.
(In response to the interlocutor’s request whether he would like to
buy any tents.)
(3) Van tandartsen word ik altijd eigenlijk helemaal nooit goed.
I always actually completely never feel good around dentists
(Following the speaker’s remark that he has a dentist appointment
next Monday)
4
Figure 1: Waveform and spectrogram of an unreduced token of eigenlijk,
produced as /ɛɪxələk/ by speaker S in the Ernestus Corpus of Spontaneous Dutch
(Ernestus 2000). The white line indicates the F2 trajectory.
Within this broad function, several subfunctions can be distinguished. As
described above, Plug (2005) investigated phonetic characteristics of 49 tokens
representing two of these subfunctions. In the present paper, we will not
distinguish between these subfunctions because they are often difficult to
distinguish and because a focus on one or several of the subfunctions would
severely reduce the number of word tokens that can be analyzed (as in Plug's
study).
The Dutch word eigenlijk can occur in different positions in the sentence
as illustrated in examples (1, 2, 3, 4, 5). Moreover, it can follow and precede
different word types, as illustrated in these same examples. Noteworthy is
example (3), in which eigenlijk is surrounded by other adverbs, as is frequently
the case in spontaneous conversations.
(4) Eigenlijk is dat actief.
Actually that is active.
(5) Het gaf heel veel informatie eigenlijk.
It gave a lot of information actually.
5
The first part of this study provides an overview of the types of variation
that we attest in our data set, and of the frequencies with which specific
phonetic characteristics occur. The second part investigates which factors predict
certain phonetic properties, including the number of syllables, presence of
creaky voice, and the presence versus absence of /l/. Our analyses will include
predictors that have been reported before to correlate with reduction degree
(e.g. speech rate, the predictability of the preceding and following word, and the
presence of sentential accent).
We also investigate the influence of a new predictor, the rhythm of the
sentence. Because the word eigenlijk can be preceded and followed by a wide
variety of words, the numbers of directly preceding and following unstressed
syllables can vary as well. Speakers of Germanic languages prefer alternating
patterns of stressed and unstressed syllables (e.g. Kelly and Bock 1988 and
references therein). It is therefore possible that speakers of Dutch opt for a
variant of eigenlijk with a number of unstressed syllables that optimizes the
rhythm of the phrase. For instance, they may prefer a stressed monosyllabic
variant if the word is followed by several unstressed syllables, and a di- or tri-
syllabic variant, ending in one or two unstressed syllables, respectively, when
the word is followed by a word with initial stress.
The next sections describe the corpus and our selection of the tokens
(Section 2), the annotation system (Section 3), and provide a qualitative
description of the attested variation (Section 4). We then present the results of
our statistical modeling of some of the tokens' characteristics (section 5). We
conclude the paper with a general discussion of these results (sections 6 and 7).
2. Materials
We extracted the tokens from the Ernestus Corpus of Spontaneous Dutch
(Ernestus 2000). This corpus, recorded in the nineties, contains high quality
recordings of ten conversations, each 90 minutes long, between pairs of friends
or direct colleagues. A DAT-recorder recorded the speech by each interlocutor on
a different track of a tape, by means of unidirectional microphones placed on a
table between the speakers. The speakers are all male, highly educated and
lived their whole lives in the Western part of the Netherlands. They speak a
'western' variant of Standard Dutch, which implies, among other things, that
they do not distinguish between the voiced and voiceless velar fricative.
The conversation during the first part of each recording was elicited by a
third person, who knew at least one of the speakers well. The speakers
discussed topics as diverse as television quizzes, how they chose their
professions, their experiences with dentists, and their opinions of euthanasia. In
the second part of the recording, the speakers participated in a role-play in
which one speaker sold tents, sleeping bags and backpacks to the other speaker,
who pretended to be a shop owner. This second part also contained
conversations covering a wide range of other topics since the speakers were
6
encouraged to do so before and after the negotiations. The speech in the corpus
sounds natural and casual, as is evidenced, among other things, by ratings of six
other native speakers, the high frequency of discourse markers (including
eigenlijk), and the amount of gossip in the corpus.
The 20 speakers in the corpus produced in total 339 eigenlijk tokens. The
number of tokens per speaker ranges from 6 to 45 (see Table 1). We randomly
selected 159 tokens, taking into account the following constraints and
preferences. First, we only incorporated tokens that were produced fluently,
without laughing and without much background noise (including the
interlocutor's speech), so that detailed phonetic analysis is possible. Secondly,
we wished to have minimally five tokens per speaker, so that we could study
intraspeaker variation, and maximally eleven tokens, so that the data set would
not be dominated by just a few speakers. Third, we preferred tokens produced in
the free conversations over tokens produced in the role play and we discarded
tokens that were produced in the first ten minutes of a recording because the
speaker might not yet have been speaking very naturally. Many tokens did not
meet all these requirements and preferences and, as a consequence, we lost
Speaker G. We also decided not to incorporate Speaker K because this speaker
often stumbled over his words. Table 1 shows the resulting number of tokens
per speaker.
Table 1. The number of tokens produced by each speaker, the number
incorporated in our analyses, and the number of studied tokens that were
monosyllabic (see the section on Individual speaker differences).
Speaker ID Total tokens Tokens studied Monosyllabic tokens
(percentage of the
tokens studied)
A 6 5 5 (100%)
B 10 9 2 (22%) E 12 8 1 (13%)
F 26 10 0 (0%)
G 8 - -
H 20 8 4 (50%)
I 15 9 1 (11%) J 10 9 1 (11%)
K 28 - -
L 45 11 0 (0%)
M 20 9 7 (78%)
N 13 7 2 (29%) O 9 9 8 (89%)
P 12 11 2 (20%)
Q 26 10 1 (10%)
R 15 7 5 (71%)
S 17 10 3 (30%) T 15 9 3 (67%)
U 22 9 7 (78%)
V 10 9 5 (56%)
7
3. Labeling procedure
A phonemic transcription of the entire intonational phrase containing eigenlijk
was made by the first author, and checked by the second author. For the token
of eigenlijk, an allophonic transcription was also made, specifying voicing of /k/
and /x/, but no other detail. The number of syllables in eigenlijk was identified.
Then, labelling of prosody and of segmental detail were carried out by the two
authors independently. All cases where the transcribers used different labels
were resolved by joint listening, as were all cases where they placed labels more
than 20 ms apart, which was not often the case. There was no obvious bias
towards either transcriber’s labelling.
For the prosodic labelling, the boundaries of the intonational phrase
containing eigenlijk were annotated, and all syllables within this phrase were
labelled as primary-accented, secondary-accented, stressed, or unstressed, that
is, we distinguished four levels of prosodic strength, defined as follows.
Primary-accented: the most prominent pitch-accented syllable in the
phrase. All phrases included minimally one primary-accented syllable; only six
included two primary-accented syllables, and most of these were produced by
the same speaker and contained equally prominent accents on eigenlijk and
another word.
Secondary-accented: lexically-stressed syllables that were produced with
a pitch movement or, rarely, a substantial increase in loudness in the absence of
a pitch accent.
Stressed syllables: lexically-stressed syllables that were produced without
a pitch movement. Unaccented function words were labelled as stressed
unaccented if they had a full vowel and no evidence of segmental reduction, or
as unstressed otherwise.
Unstressed: syllables lacking lexical stress. Filled pauses were always
labelled as unstressed.
For the labelling of segmental detail, we defined a set of articulatory
events in the larynx and supraglottal tract which are present in an unreduced
token of eigenlijk, including the onsets and offsets of periodicity and the onsets
and offsets of creaky voice. These events are described in detail in the following
paragraph and illustrated in Figure 2. If an event was identifiable in the
waveform and spectrogram, it was labelled. If it was absent or unidentifiable
because of reduction, then that label was omitted. By labelling events rather
than segments, we aimed to achieve maximum comparability across tokens that
had different degrees of reduction, while avoiding parsing the signal exhaustively
into phoneme-sized segments, which can be very challenging for reduced
speech. We made one exception: if a velar stop was present, we marked its
offset in the signal, even if the stop was unreleased and directly followed by
another stop. In these cases we placed the boundary for the stop in the middle
of the long closure formed by the two stops. We thus maximized the number of
8
velar stops whose durations we could analyse (see below). However, utterance-
final unreleased stops were excluded from this analysis, as we had no principled
way to estimate their durations.
As Figure 2 shows, on the larynx tier we labelled the onsets and offsets of
periodicity (numbered as P1 and xP1 for the first portion of periodicity, P2 and
xP2 for the next, etc). We also labelled the onset (CRK) and offset (xCRK) of
creaky voice, if present during /ɛɪ/. Our criterion for identifying creaky voice was
irregularity of periods. On the upper articulators tier we labelled the following
acoustic events: the onset of the stressed /ɛɪ/ vowel (Vo); the onset and offset of
velar frication (VF, xVF respectively), of lateral quality (L, xL) and of velar
closure (V, xV). These labels allowed us to calculate the duration of the whole
word token (defined as extending from the onset of the stressed vowel to the
last labelled event) and durations of individual segments, including the
unstressed vowels, if present.
9
Figure 2: Labelling of segmental detail of two tokens of eigenlijk. Top, an
unreduced token produced as /ɛɪxələk/ by speaker H. Bottom, a reduced token
produced as [ɛɪg] (phonemically /ɛɪk/) by speaker M (the following word, niet
‘not’, is also shown). In each case, the top tier indicates laryngeal events, the
second tier indicates events involving the upper articulators, and the third tier
shows the allophonic transcription (see text for details). The white line indicates
the F2 trajectory.
4. General description of variation within the word
The Appendix lists the variation that we discuss in this and the following section
for which we can provide frequency data.
We first focus on the variation in the number of syllables. Figure 3 shows
the number of mono-, di- and trisyllabic tokens for the four prosodic statuses
that we distinguished. We see that the majority of tokens are disyllabic, and are
thus one syllable shorter than the full form. Moreover, we find that many
disyllabic and some monosyllabic tokens are accented (8 monosyllabic tokens
are primary-accented and 16 secondary-accented). The word eigenlijk thus does
not follow the well-known pattern that accented word tokens show little
10
reduction (e.g. Bell et al. 2003). The variation in the number of syllables does
not just result from the prosodic status of the word.
Figure 3: Number of trisyllabic, disyllabic and monosyllabic tokens that are
unstressed, stressed, secondary accented or primary accented.
With regard to duration, tokens were on average longer when they had a greater
number of syllables (3 syllables: 386 ms; 2 syllables: 310 ms; 1 syllable: 197
ms), but Figure 4 shows that the durational ranges for tri-, di- and monosyllabic
tokens overlapped considerably, and we observed trisyllabic tokens as short as
257 ms.
Figure 4: Boxplot of the duration of eigenlijk (ms), according to its number of
syllables. The bottom and top of each box indicate the first and third quartiles;
the band inside the box represents the second quartile (the median), while the
0
10
20
30
40
50
60
70
80
unstressed stressed secondaryaccented
primaryaccented
monosyllabic
disyllabic
trisyllabic
Number
of tokens
11
whiskers extend to the minimum and maximum values that are maximally 1.5
interquartiles from the box. The two small circles are outliers.
Table 2 lists the phonemic transcriptions of the tokens, and allophonic
transcriptions specified for the voicing of /k/ and /x/ but showing no other detail,
along with the frequency of each transcription. The most frequent phonemic
form is disyllabic /ɛɪxlək/ (57 tokens). The monosyllabic form /ɛɪk/ (22 tokens)
comes in second position, and the full form /ɛɪxələk/ (20 tokens) in third. Also
common are monosyllabic forms /ɛɪxk/ (16 tokens) and /ɛɪx/ (13 tokens) and
the disyllabic /ɛɪxək/ (14 tokens).
Importantly, we did not see a clear correlation between the structure of a
token and how clearly its segments were articulated. Tokens that are highly
reduced in terms of their number of syllables and segments, could nonetheless
exhibit clearly articulated and tightly coordinated segments, and vice versa, as
discussed below.
Table 2 makes clear that all forms in our dataset contain, minimally, a full
front vowel, and at least one obstruent articulated at the back of the mouth.
These appear to be essential phonetic components of eigenlijk. The vowel is
typically a closing diphthong, but varies in degree of diphthongization, and can
look and sound quite monophthongal (e.g. Figure 2, right panel; Figure 9). As
regards the obstruent(s), 114 tokens (72%) were transcribed as containing both
a fricative and a stop, 19 (12%) only a fricative and 26 (16%) only a stop. The
obstruents are typically voiceless, but are voiced in 18% of cases. The fricative’s
place of articulation is normally velar or uvular. Eighty-seven tokens (55%) were
transcribed as containing /l/, while several more contain a residual trace of /l/,
as discussed further below.
Table 2. Tokens of eigenlijk found in the corpus. Phonemic and allophonic
transcriptions are shown. The allophonic transcriptions specify the voicing of /k/
and /x/, but no other detail. Note that [ɛɪxl] was once perceived as disyllabic and
once as monosyllabic.
12
Token structure N Transcription
Phonemic Allophonic
Trisyllabic 20
Vowel + fricative + schwa +
lateral + schwa + stop
20 /ɛɪxələk/ 20 [ɛɪxələk] 17, [ɛɪxələg] 2,
[ɛɪɣələk] 1
Disyllabic 82
Vowel + fricative + lateral + schwa + stop
60 /ɛɪxlək/ 57
/ɛɪxləŋ/ 1
/ɛxlək/ 2
[ɛɪxlək] 42, [ɛɪxləg] 7,
[ɛɪɣlək] 8, [ɛɪɣləg] 1
[ɛɪxləŋ] 1
[ɛxlək] 1, [ɛxləg] 1
Vowel + fricative + schwa + stop
15 /ɛɪxək/ 14
/ɛxək/ 1
[ɛɪxək] 6, [ɛɪxəg] 2, [ɛɪɣək] 5,
[ɛɪɣəg] 1
[ɛxəg] 1
Vowel + fricative + lateral 1 /ɛɪxl/ 1 [ɛɪxl] 1
Vowel + fricative + schwa +
lateral
1 /ɛɪxəl/ 1 [ɛɪxəl] 1
Vowel + lateral + schwa + stop 4 /ɛɪlək/ 4 [ɛɪlək] 3, [ɛɪləg] 1
Monosyllabic 57
Vowel + fricative 16 /ɛɪx/ 13
/ɛx/ 3
[ɛɪx] 11, [ɛɪɣ] 2
[ɛx] 2, [ɛɣ] 1
Vowel + stop 22 /ɛɪk/ 22 [ɛɪk] 13, [ɛɪg] 9
Vowel + fricative + stop 18 /ɛɪxk/ 16
/ɛxk/ 2
[ɛɪxk] 15, [ɛɪɣk] 1
[ɛxk] 2
Vowel + fricative + lateral 1 /ɛɪxl/ 1 [ɛɪxl] 1
Total 159
Trisyllabic tokens were produced as the canonical form /ɛɪxələk/. Figures 1 and 2
(left panel) show typical trisyllabic tokens. Note the formant dynamics, in
particular how F2 rises through the first diphthong to reach a maximum at the
start of /x/, then falls to reach its minimum during /l/, before rising again into
/k/. The first schwa is typically shorter than the second (mean durations 19
versus 46 ms, respectively).
Despite containing all or almost all of the segments expected in the
canonical form, some of the trisyllabic tokens did display elements of reduction.
Some were quiet and/or breathy, particularly if phrase-final. Others were rapidly
articulated: Figure 5 shows a trisyllabic token that is very short indeed (268 ms)
and whose formant dynamics follow a less extreme version of the pattern
described above. Among the types of reduction found in trisyllabic tokens were
also monophthongisation of the vowel, incomplete closure of the stop, or
devoicing of the first schwa.
13
Figure 5: Trisyllabic token of eigenlijk produced as /ɛɪxələk/ by speaker E. Note
the short duration and the less extreme excursions of F2 (indicated by the white
line) compared to the trisyllabic tokens in Figures 1 and 2. The white line
indicates the F2 trajectory.
Disyllabic tokens took a wide range of forms. The most common disyllabic
form was /ɛɪxlək/, very similar to trisyllabic tokens, but with no discernible
schwa before /l/. Such schwa loss in unstressed syllables before /l/ and /r/ is
very common also in English (e.g. Gimson & Cruttenden 1994).
Disyllabic tokens evinced a range of interesting phonetic behavior at the
juncture between the stem and suffix. First, there was variation in the extent to
which the expected laryngeal events were produced, and in how they were
aligned with respect to events involving the upper articulators. For example, we
encountered numerous cases of progressive voice assimilation of /l/ to the velar
fricative, that is, cases where /l/ and on occasion the word’s entire second
syllable was devoiced (e.g. Figure 6). We also found cases of regressive voice
assimilation, that is, where /x/ was voiced by assimilating to /l/, sometimes
resulting in a token that was voiced in its entirety (e.g. Figure 7). Voice
assimilation involving /l/ has not been described for Dutch so far. When voicing
of the fricative occurred, it was sometimes accompanied by weakening of the
degree of stricture, and/or loss of place cues, such that the frication sounded
glottal rather than velar or uvular. In four disyllabic tokens, phonemically /ɛɪlək/,
14
we found no evidence of velar frication at all, but simply a lateral approximant at
the syllable boundary.
Figure 6: Disyllabic tokens of eigenlijk produced as /ɛɪxlək/, illustrating devoicing
in the second syllable. Left: Token produced by speaker N, with a devoiced /l/.
Right: Token produced by speaker V, where the second syllable is devoiced, but
preserves the formant dynamics consistent with a /lə/ sequence. Dashed lines
on spectrograms indicate the boundaries of /l/. White lines indicate the F2
trajectory.
15
Figure 7: Fully voiced disyllabic token of eigenlijk produced as [ɛɪɣləg]
(phonemically /ɛɪxlək/) by speaker T. The white line indicates the F2 trajectory.
Second, in disyllabic tokens we found variation in the alignment of events
involving the upper articulators. In several tokens, most of them spoken by
speaker F, the /x/ and /l/ appear strongly coarticulated. This speaker seems to
produce lateral frication (e.g. Figure 8, left panel) as a solution to the problem of
producing two very different articulations in swift succession. A different strategy
is adopted by speaker S (e.g. Figure 8, right panel), who appears to start
backing the tongue body in preparation for the /l/ already from the start of the
frication, such that F2 reaches its minimum in the middle of the fricative portion.
Finally, among the tokens transcribed phonemically as /ɛɪxək/, we also observed
cases where an /l/ was not unambiguously present, but nevertheless left some
residual trace in the signal, in the form of an F2 dip (e.g. Figure 9).
16
Figure 8: Disyllabic tokens of eigenlijk in which strong coarticulation between
velar frication and laterality is audible. White lines indicate F2 trajectories. Left
panel: token produced as /ɛɪxlək/ by speaker F, i.e. with an apparent lateral
fricative [ɬ] in place of /xl/. Right panel: token produced as /ɛɪxlək/ by speaker
S, in which F2 (indicated by the white line) falls rather steeply from the start of
the frication.
17
Figure 9: Disyllabic token of eigenlijk, produced as [ɛxəg] (phonemically /ɛxək/)
by speaker I in the context dat doe ik eigenlijk nooit 'I actually never do that'.
This token was not heard as containing a definite /l/, but the spectrogram
indicates a residual trace of /l/, manifest as an F2 dip around 0.15 seconds (the
white line indicates F2, and the black dotted line the F2 minimum). Note also the
assimilation of the final stop to the following /n/ in terms of voicing and nasality.
The black dashed line indicates the start of nooit (produced with laughter).
Monosyllabic tokens also took a number of forms. Some ended in a sequence of
fricative followed by stop (phonemically /ɛɪxk/). The obstruent cluster /xk/ is not
a legitimate syllable coda in Dutch. It was often produced with rather long
duration relative to the vowel (e.g. Figure 10). A small number of tokens, with
particularly long obstruent clusters, were difficult to classify in terms of their
number of syllables: despite having only one syllable peak, they sounded almost
disyllabic (see Aoyagi in press for a possible theoretical account of this finding).
18
Figure 10: Monosyllabic tokens of eigenlijk, produced as /ɛxk/ by speaker I (left
panel) and as /ɛɪxk/ by speaker O (right panel). Note the long duration of the
obstruent portion compared to the vowel in both tokens. White lines indicate F2
trajectories.
Other monosyllabic tokens contain a single voiceless obstruent, either /x/ or /k/.
The formant dynamics of the vowel are quite variable in these tokens. Some
tokens show a flat or falling F2 during the vowel (e.g. Figure 11). Others have a
typically diphthongal vowel ending in a clear velar pinch (convergence of F2 and
F3) at the transition into the obstruent (e.g. Figures 12 and 13). We are not sure
to what to attribute the difference between these two patterns, but it may relate
to the place of articulation of the obstruent: the diphthongal vowels may precede
consonants with a velar place, and the vowels with flat or falling F2 may precede
post-velar or uvular consonants (for comments on how a velar vs. uvular place
distinction can affect vowel formants, see Gordon, Barthmaier and Sands 2002).
19
Figure 11: Monosyllabic tokens of eigenlijk, produced as /ɛx/ by speakers P (left
panel) and N (right panel). In each case F2 is indicated by a white line; note the
flat or falling F2 at the transition into the obstruent. White lines indicate F2
trajectories.
20
Figure 12: Monosyllabic tokens of eigenlijk, produced as /ɛɪk/ by speakers S (left
panel) and R (right panel). F2 is indicated in each case by a white line; note its
rise and the convergence of F2 and F3 (velar pinch) at the transition into the
obstruent. White lines indicate F2 trajectories.
21
Figure 13: Monosyllabic token of eigenlijk, produced by speaker S as [ɛɪg]
(phonemically /ɛɪk/) in the context eigenlijk door 'actually by'. Note the velar
pinch at the transition into the stop. The white line indicates the F2
trajectory.The final stop is also assimilated in voice to the following /d/, and is
unreleased.
Finally, creaky voice was common in the initial full vowel /ɛɪ/. This vowel often
carries stress or accent (see e.g. Figure 3), and according to van Jongenburger
and van Heuven (1991), the word may therefore be expected to be preceded by
a glottal stop or similar phonetic events (i.e., glottalisation/creaky voice),
especially after a vowel. Indeed, half of our tokens showed creaky voice at the
start of the vowel (83 cases, or 52% of the data set), and less frequently other
types of non-modal voice quality, such as harshness or breathiness.
4.2 Voice assimilation
The eigenlijk tokens, whether tri-, di- or mono-syllabic, show unexpected
patterns of voice assimilation. Dutch is typically assumed to have only two
processes of voice assimilation affecting sequences of obstruents. The first
process voices obstruents preceding /b/ and /d/, while the second devoices /v/
and /z/ after voiceless obstruents (e.g. Booij 1995: 58, 59). We observed
examples of these processes: see e.g. Figure 13 above. Yet we also found cases
22
of assimilation not described in the literature. Thirteen tokens showed voicing of
the word-final velar obstruent before nasal-initial words such as niet 'not', nooit
'never', and nog 'yet' (e.g. Figure 2, right panel, which illustrates an assimilated
final stop in a monosyllabic token of eigenlijk, in the set phrase ik weet het
eigenlijk niet 'I actually don’t know'; Figure 9; Figure 14) and we also observed
voicing before /ʋ/ in one case. Typically for tokens of this kind, the stop is weak
and very short, as little as 20 ms, relative to a vowel lasting 110-160 ms.
Furthermore, both /x/ and /k/ were sometimes voiced when followed by a
vowel, whereas intervocalic voice assimilation at prosodic word boundaries is
assumed to be restricted to fricatives (Booij 1995: 147). Finally, a handful of
tokens of eigenlijk were followed by devoiced nasal stops, probably resulting
from progressive voice assimilation induced by /k/ (e.g. Figure 15). All in all, the
tokens of eigenlijk in our data set show more voice assimilation than would be
expected on the basis of the existing literature.
Figure 14: Tokens of eigenlijk where the final obstruent assimilates in voice to a
following nasal segment. Left panel: [ɛɪxələg] (phonemically /ɛɪxələk/), in the
context eigenlijk maak ik… 'actually I make...', by speaker I. Right panel: [ɛɪg]
(phonemically /ɛɪk/), in the context ik weet het eigenlijk niet precies 'I actually
do not know exactly', by speaker U. White lines indicate F2 trajectories.
23
Figure 15: Token of eigenlijk produced as /ɛɪx/ by speaker O in the context
eigenlijk niet ‘actually not’. The token is followed by a partially devoiced /n/
which has assimilated in voice to the final segment of eigenlijk. The white line
indicates the F2 trajectory.
4.3 Individual speaker differences
Although the speakers form a homogeneous group (they are all adult speakers
coming from the same region and from the same socio-economic class), they
clearly show individual differences. Speakers vary in their propensity to produce
tokens with different numbers of syllables. This is illustrated in Table 1, which
lists for each speaker the percentage of monosyllabic tokens. Speaker A
produced only monosyllabic tokens, speakers F and L produced no monosyllabic
tokens, and all other speakers produced a mixture of monosyllabic and di-/tri-
syllabic tokens. There is also clear variation in the frequency of trisyllabic
tokens: half of these were produced by only three speakers (E, J, and S). Part of
this variation may be accounted for by individual differences in speech rate (see
the next section).
In addition, there may be a role for the position of the word in the
prosodic phrase. The speakers varied in terms of where in the prosodic phrase
their tokens of eigenlijk tended to occur. Speakers E, F, and J produced over half
of their tokens at phrase edges, whereas all others produced the majority of
24
their tokens phrase-medially. This variation in prosodic position probably partly
explains the interspeaker variation in reduction degree because word tokens at
prosodic boundaries tend to be less reduced (e.g. Bell et al. 2003).
At the level of phonetic detail, while most of the patterns observed were
common to more than one speaker, there were certain production strategies
that appeared to be specific to one or just a few speakers. For instance,
speakers F and L were the only ones in the dataset who produced overlapping
velar frication and /l/-quality. They contrast with speaker N, for instance, who,
often devoiced the /l/. Furthermore, impressionistically, some speakers were
very variable in their realisations of eigenlijk whereas others were more
consistent. Further research has to investigate to what extent physiological
differences between the speakers may explain these individual reduction
patterns.
5. Analysis of the conditions under which some properties appear
5.1 Predictors
We tested whether several properties of the tokens may be conditioned by the
following five types of predictors. First, we investigated the influence of the
predictability of the preceding and following word, since both have been shown
to correlate with the duration and the number of segments of eigenlijk
(Pluymaekers, Ernestus, and Baayen 2005b). We defined these predictabilities
as the logarithms of the numbers of occurrences of these words plus one in the
Spoken Dutch Corpus (ranges for both words: 0 - 12.16).
Secondly, we studied the influence of a temporal measure. As mentioned
in the Introduction, many studies have shown that a higher speech rate
generally favors a higher reduction degree. Pluymaekers, Ernestus, and Baayen
(2005b) showed that this also holds for eigenlijk. We defined speech rate as the
logarithm of the number of syllables per second in the citation forms of the
words in the labeled phrase (mean: 7.7 syllables/second; range: 1.3 - 20.5
syllables/second). We used the number of syllables of the citation form because
it is an important predictor of perceived rate (Koreman 2006) though the
realized number of syllables also plays a role. We applied a logarithmic
transformation because an increase of one syllable per second is likely to have a
bigger impact if the rate is one syllable per second than if it is seven syllables
per second.
Thirdly, we included prosodic measures: the level of accent on eigenlijk
and its position in the phrase. We started with regression models in which both
predictors had four levels (primary accented, secondary accented, stressed, and
unstressed; isolation, phrase-initial position, phrase-final position, phrase-medial
position), but models with these predictors did not converge. We therefore
25
simplified these predictors to two-level predictors (accented or not; in phrase-
medial position or not).
Further, we included as prosodic measures the number of unstressed
syllables preceding eigenlijk and the number of unstressed syllables following
eigenlijk (both ranges: 0 - 3). They provide information about the rhythm of the
phrase. We compared these two continuous measures with two predictors that
merely indicate whether the preceding/following syllable is unstressed. The
tables with the statistical results show the models with these categorical
variables. If these models differ significantly from the models with the
continuous variables, this is mentioned in the text.
Fourthly, we investigated the roles of the types of preceding and following
segments, since through co-articulation they may directly affect the realization
of the neighboring segments. These variables, however, never played
statistically significant roles.
Finally, we also tested whether the number of syllables could predict the
other properties of the tokens. Because we feel uncomfortable modelling one
dependent variable with another one, we only report the results of these
analyses in the text (i.e. without details in tables). For the same reasons, we did
not incorporate in the main analysis vowel duration as a predictor for creaky
voice.
5.2 Statistical analyses
We analyzed the continuous dependent variables with linear mixed effects
modeling, and the Boolean dependent variables with generalized linear mixed
effects modeling with the logit link function, as implemented in the statistical
package R (R Core team, 2014). We tested for random intercepts for speaker,
preceding word and following word. We did not include random slopes because
preliminary testing suggested that models including them did not converge or
seemed to overfit the data.
Our final models, reported below, only include statistically significant
predictors. Fixed predictors were considered significant if their absolute t-values
or z-values were greater than 1.96 (which approximates an alpha level of 0.05).
Random intercepts were considered significant if the model with the random
intercept outperformed the model without that random intercept as indicated by
likelihood ratio tests (again we adopted an alpha level of 0.05). For continuous
dependent variables, the final models are only based on those data points that
differed less than 2.5 standard deviations from the values predicted by the
model.
5.3 Results
We first studied which variables predict the number of syllables of eigenlijk, by
analyzing two dependent Boolean variables: whether the word contains one
syllable or whether the word contains three syllables. Trisyllabic tokens are
relatively rare in the dataset (only 20 tokens) and it is therefore not surprising
26
that there are fewer predictors of the occurrence of trisyllabic than of
monosyllabic tokens (see Table 3).
We found that monosyllabic tokens are more likely at a higher speech rate
and when the preceding word is of a higher frequency of occurrence. In addition,
polysyllabic tokens are followed by stressed syllables in 71% of cases, whereas
monosyllabic tokens are approximately equally often followed by unstressed and
stressed syllables (51% versus 49%). This suggests a relatively strong tendency
for eigenlijk to be monosyllabic when followed by an unstressed syllable. The
categorical variable, which only provides information about whether the next
syllable is stressed, results in a model that is as good as the model with the
continuous variable indicating the exact number of following unstressed syllables
(the two variables results in models with similar Akaike information criterion
values: 152 versus 151, Akaike 1974).
Trisyllabic tokens mostly occurred at lower speech rates and at phrase
boundaries. Out of the 20 trisyllabic tokens, only five occurred in phrase-medial
position (20%), whereas no less than 105 out of the 139 mono- and disyllabic
tokens (76%) occurred phrase-medially.
As expected given our previous observations on differences between
speakers (see section 4.3), speaker is an important random effect. In addition,
the probability of a monosyllabic token was conditioned by the identity of the
following word. We also found random effects of speaker and following word in
the analyses presented below, and we will no longer mention them separately.
Table 3: Statistical results for the number of syllables. A positive coefficient
implies that the predictor increases the probability of one/three syllable(s).
Fixed effects Exactly one syllable Exactly three syllables
Coefficient z-value Coefficient z-value
Intercept -11.963 -3.48 3.597 1.81
Speech rate 4.192 2.81 -2.524 -2.53
Preceding word
frequency
0.358 2.37 - -
No following unstressed
syllables
-1.996 -2.00 - -
Phrase-medial position - - -2.311 -3.49
Random effect Variance SD Variance SD
Speaker 6.965 2.64 1.712 1.31
Following word 4.646 2.16 - -
We then analyzed the duration of the whole word token, and of the most
frequently occurring parts. Table 4 shows the statistical results for the duration
of the token as a whole and that of its first, full, vowel. As expected, both units
are shorter at a higher speech rate and if non-accented (mean duration of non-
accented tokens: 267 ms; mean duration of accented tokens: 294 ms). The
complete token is also shorter when in phrase-medial position (mean duration:
255 ms) than at phrase boundaries (mean duration: 337 ms).
27
The duration of the whole token is in addition affected by the rhythm of
the phrase: tokens tend to be shorter (mean of 259 ms versus a mean of 293
ms) if they are followed by (any number of) unstressed syllables. This result is in
line with the results for the probability of a monosyllabic variant, presented
above. This effect of the rhythm of the phrase only surfaces in the analysis of
token duration if we test the categorical variable. A variable that exactly
indicates how many unstressed syllables are following is not predictive of the
duration of the token, and results in a model with a slightly higher Akaike
Information Criterion value (1777 versus 1775).
The number of syllables is a good predictor for the durations of both the
complete token and the full vowel. The statistical results for vowel duration
hardly change if the number of syllables is incorporated as an additional
predictor. This is different for the token duration. The effect of the rhythm of the
phrase on the duration of the token is no longer significant, which is not
surprising because the number of syllables of eigenlijk and the rhythm of the
phrase are correlated (see above). In addition, the duration of the token is no
longer predicted by whether the token is accented or not. This may be more
surprising because accentedness does not predict number of syllables in our
analyses presented above. Possibly, these analyses were not sufficiently
sensitive since we investigated two Boolean dependent variables (monosyllabic
or not; trisyllabic or not).
Table 4: Statistical results for word and full vowel durations
Predictor Word duration Vowel duration
Coefficient t-value Coefficient t-value
Intercept 525.115 12.68 229.666 10.47
Speech rate -112.823 -5.66 -41.141 -3.89 Phrase-medial position -35.455 -2.86 - - Non-accented -24.350 -2.67 -18.650 -3.79
No following unstressed
syllable
24.603 2.37 - -
Random effect Variance SD Variance SD
Speaker 512.9 22.65 109.7 10.48
Following word 1606.0 40.08 - -
Table 5 shows the statistical results for the duration of the velar fricative, if it
was present (133 tokens). The fricative was shorter at higher speech rates and
when in phrase-medial position (mean duration: 61 ms) rather than at a phrase
boundary (mean duration: 68 ms). These variables also predicted token or vowel
durations, and in the same directions (see above). The token's number of
syllables does not predict the duration of the fricative.
If we test the effect of the rhythm of the phrase on the duration of the
fricative with the categorical variables, we found no effects (as indicated in Table
5). We find an effect, in contrast, if we test the continuous variable indicating
the exact number of following unstressed syllables (coefficient: 6.705; t-value:
2.58). Surprisingly, the fricative tends to be longer if eigenlijk is followed by a
28
higher number of following unstressed syllables. The absence of an effect of the
categorical variable and the unexpected direction of the effect of the continuous
variable (which is also opposite to what we found for the probability of a
monosyllabic form and for the duration of the complete token) raises the
question whether this effect on the fricative duration may be a Type 1 error or
just arises because the number of following unstressed syllables happens to be
correlated with some other relevant predictor.
Each of the other segments (i.e. the segments in the suffix -/lәk/) was too
often absent to allow for an analysis of its duration. We therefore analyzed the
duration of the suffix as a whole, in the 140 tokens in which at least one of its
segments was present and in which the final segment was not an unreleased
stop followed by silence (see section 3). The results are presented in the last two
columns of Table 5. The suffix is shorter at higher speech rates and if in phrase-
medial position rather than at phrase boundaries (mean durations: 83 ms versus
135 ms). The effect of speech rate disappears if the number of syllables is taken
into account. As expected, the suffix is longer if the token contains more
syllables and the suffix is thus present and does not only consist of a velar
consonant.
Table 5: Statistical results for the durations of the velar fricative and the suffix -
/lәk/
Predictor Fricative duration Suffix duration
Coefficient t-value Coefficient t-value
Intercept 100.916 8.18 192.672 7.98
Speech rate -17.372 -2.86 -33.653 -2.83
Phrase-medial position -7.867 -2.34 -39.728 -5.18
Random effect Variance SD Variance SD
Speaker 18.01 4.24 159.2 12.62
Following word - - 474.6 21.79
From all segments, only /l/ was both present in many tokens (88) and absent in
many tokens (71). This was therefore the only segment for which we could
analyse which variables predicted its presence. The results are presented in
Table 6. The segment /l/ was less often present at higher speech rates and when
located in phrase-medial position (47% of tokens) rather than at a phrase
boundary (84%). Moreover, /l/ was more often realized if the eigenlijk token
was not followed by unstressed syllables (63%) than if one or more unstressed
syllables followed (39%). This latter effect of rhythm only surfaces if we test the
categorical rather than the continuous variable indicating the exact number of
following unstressed syllables.
All effects disappear if the number of syllables in the spoken eigenlijk
token is taken into account. The strong effect of the number of syllables is
unsurprising: /l/ is seldom present in monosyllabic tokens, quite often present in
disyllabic tokens, and always present in trisyllabic tokens (see Table 2). The
number of syllables is predicted by the same variables as the presence of /l/
(see Table 3).
29
Table 6: Statistical results for the presence of /l/ and of creaky voice
Predictor /l/ Creaky voice
Coefficient z-value Coefficient z-value
Intercept 6.1744 -2.76 3.098 1.87
Speech rate -2.6348 -2.49 -1.715 -2.16
Phrase-medial position -1.8897 -3.29 -0.920 -2.10
No following unstressed syllable 1.1174 2.22 Following word frequency - - 0.128 2.23
Random effect Variance SD Variance SD
Speaker 2.746 1.66 0.968 0.98
Finally, we analyzed creaky voice, a variable that has not been analyzed so far
as a measure of degree of reduction. Creaky voice was present in 83 tokens. We
modeled the probability that creaky voice was present as well as the duration of
creaky voice if present. The results for the presence of creaky voice are
presented in the last two columns of Table 6. Creaky voice was less often
present at higher speech rates and when the token was in phrase-medial
position (in 52% of phrase-medial tokens) rather than in phrase-initial or -final
position (in 63% of these tokens). Furthermore, there was a correlation with the
frequency of the following word: a more predictable following word appears to
increase the likelihood of creaky voice.
The presence of creaky voice cannot be predicted by the number of
syllables of the token of eigenlijk. In contrast, an additional analysis established
that the presence of creaky voice can be predicted by the duration of the vowel.
(Recall that we did not include vowel duration in the main analysis for creaky
voice because we did not want to model one dependent variable with another
one; see 5.1 above.) If vowel duration is incorporated as a predictor in the
model for the presence of creaky voice, the effect of the following word
frequency is still statistically significant, while those of speech rate and of phrase
medial position are only marginally significant (ps < 0.07).
The duration of creaky voice, if present, could not be predicted by any of
our independent variables. We could only find a high correlation with vowel
duration (coefficient: 0.228; t = 2.02), showing that the longer the vowel, the
longer the part with creaky voice tends to be. In addition, we found a random
effect of speaker (variance: 281; S.D. = 16.76).
6. General Discussion
This chapter has presented a detailed analysis of the properties of 159 tokens of
the Dutch discourse marker eigenlijk, which occur in the Ernestus Corpus of
Spontaneous Dutch (Ernestus 2000). Our aim was to document the wide
variation in the pronunciation of the word and to analyze which properties of the
context predict this variation. Previous research suggests that the exact meaning
of a token has an influence on its detailed phonetic characteristics (Plug 2005).
We did not distinguish between the different meanings because this would have
30
resulted in too small a data set for statistical analyses. Moreover, the meaning of
the word only seems to favor some pronunciation variants rather than
completely excluding others. As such, it would have only been one of our many
predictors. It would be worthwhile investigating the role of the word's exact
meaning in a larger data set.
The qualitative analysis of the tokens in our dataset supported previous
studies (e.g. Ernestus 2000) in demonstrating a wide range of variation in the
production of the word, ranging from trisyllabic tokens closely resembling the
word’s citation form, through to phonetically minimal monosyllabic tokens
consisting merely of a vowel followed by a single obstruent consonant. The
disyllabic forms occur most frequently (52% of tokens), followed by the
monosyllabic variants (36%), while trisyllabic forms are relatively rare (13%).
Reduction of eigenlijk may be manifest in a number of ways, in the
number of syllables that a token contains, or in its duration, or the clarity of its
articulation. We expected these properties to correlate with one another. In fact,
we found surprisingly little clear correlation between the different indices of
reduction. Although the number of syllables in a token did increase along with
duration, the durational ranges found in mono-, di- and tri-syllabic tokens clearly
overlapped. Moreover, some tokens that had only one syllable nonetheless had
clearly articulated and tightly coordinated segments, while some di- and
trisyllabic tokens appeared to be articulated rather laxly. These observations
underscore that reduction is not a simple or automatic consequence of speaking
under time pressure.
We found some support for the concept of “articulatory prosodies” in the
sense of non-linear features of syllables that are no longer tied to specific
segmental units (Niebuhr and Kohler, 2011). In particular, the /l/ of eigenlijk is
not always present, but when a definite lateral articulation is absent, acoustic
and auditory traces of /l/ often remain. Also, the weak syllables of the word may
be absent yet leave an acoustic residue in the signal, for instance, as extra
duration of the consonants in monosyllabic tokens produced as /ɛɪxk/. Following
Niebuhr and Kohler’s (2011) reasoning further, can we specify a “phonetic
essence” of the word eigenlijk? Apparently, the only essential components are a
front, usually diphthongal vowel and at least one back (velar or uvular)
obstruent. Even these essential parts allow for variation: the vowel can lose its
diphthongal quality and can become somewhat more backed; the obstruent can
be either stop or fricative, and given an appropriate conditioning context, it can
lose its voicelessness.
We observed unexpected patterns of voice assimilation within the word
and at word boundaries that are not described in the literature. We found both
regressive and progressive voice assimilation of /l/ within the word, and of the
final voiceless obstruents of eigenlijk to following nasal consonants. Local (2003)
proposes that different patterns of assimilation occur for function words (e.g.
I’m) compared to content words (e.g. lime). Our data suggest that the same
may be true for discourse markers, and perhaps for frequent sequences such as
31
eigenlijk niet 'actually not'. Further work is needed to show whether the
observed assimilation patterns are indeed specific for eigenlijk.
Also unexpectedly, the dataset showed no clear relationship between a
token's prosodic status and its reduction in terms of either duration or number of
syllables. We observed a surprisingly large number of accented tokens that were
produced with only one syllable (eight primary-accented and 16 secondary-
accented monosyllabic tokens). This clearly shows that reduction of eigenlijk is
not a phenomenon restricted to prosodically weak positions. A token may be
heavily reduced in duration or number of syllables, yet may still constitute the
most prominent word in its local context. Future studies have to show which
other (types of) words can also be drastically reduced in prosodically strong
positions.
The dataset contains tokens from 18 speakers from a rather
homogeneous group (all highly educated men raised in the Western part of the
Netherlands). Nevertheless, there are substantial differences between speakers
in their reduction degree. Some speakers are clearly more likely to produce
monosyllabic forms of eigenlijk than others. Speakers also differ in how they
solve the problem of quickly producing a velar/uvular fricative followed by a
lateral, for instance by complete coarticulation of the two sounds, resulting in a
lateral fricative, or by weakening of the degree of stricture for the fricative.
Finally, in our analyses investigating which variables predict the duration and
presence versus absence of segments, speaker was always a significant random
effect.
These individual patterns confirm and extend the results by Hanique,
Ernestus and Boves (2015). By means of computational modeling of
automatically generated segmental transcriptions of the speech in the entire
corpus, these researchers showed that the speakers in our data set can be
better distinguished from each other if not only the word types and combinations
of word types that they produced are taken into account but also how these
speakers reduced phones and combinations of phones. Our study contributes to
the results by Hanique and colleagues by documenting interspeaker variation in
the pronunciation at the segmental level of one entire word (instead of a single
phone or a sequence of three phones), which forms part of the information the
computer modelling of Hanqiue and colleagues was based on. On top of this, our
results show that the speakers differ at the subsegmental level, which was not
taken into account by Hanique and colleagues.
In the second part of the chapter, we investigated which variables may
predict the number of syllables of a token, the durations of a token and its parts,
and the presence of /l/ and creaky voice. In the Introduction to this chapter, we
hypothesized that the rhythm of the sentence may have an effect on reduction
degree of eigenlijk. Speakers of Germanic languages prefer sentences with
approximately equally long intervals between stressed syllables. In order to
minimize the sequence of unstressed syllables, they may realize tokens of
eigenlijk followed by unstressed syllables as (stressed) monosyllabic variants.
Conversely, in order to avoid stress clashes, they may prefer a di- or tri-syllabic
32
variant, ending in one or two unstressed syllables, respectively, when the word
token is followed by a word with initial stress. Our data set provides support for
this hypothesis. Tokens of eigenlijk appear to be more often monosyllabic, to be
shorter in duration, and to be less likely to contain /l/s when followed by
unstressed syllables. To our knowledge, effects of rhythm on speech reduction
have not been documented before.
Interestingly, the categorical variable that just indicates whether the
token of eigenlijk was followed by either a stressed or an unstressed syllable
outperformed a continuous variable indicating the exact number of following
unstressed syllables for token duration and for presence of /l/. (For the
probability of a monosyllabic form, the categorical and the continuous variables
are equally predictive). This suggests that only the prosodic status of the
immediately following syllable is relevant. Possibly, when producing eigenlijk,
speakers have only taken decisions about the prosodic status of the next
syllable. Another possible explanation is that speakers cannot or are not inclined
to reduce eigenlijk even more when it is followed by more than one unstressed
syllable.
Unexpectedly, it is the exact number of following unstressed syllables
rather than the presence of a following unstressed syllable that predicts the
duration of the fricative. Moreover, this effect goes in the non-hypothesized
direction: this fricative was longer if it was followed by more unstressed
syllables. We do not know how to explain this unexpected result.
Of the fixed predictors that have been shown before to correlate with
reduction degree, two emerged as statistically significant in (nearly) all
analyses: tokens were more reduced at higher speech rates and in phrase-
medial position. These results replicate previous findings (e.g. Bell et al. 2003;
Kohler 1990, 2009; Raymond, Dautricourt, and Hume 2006). We found these
effects also for the probability of creaky voice, that is, creaky voice was less
likely to occur at higher speech rates and phrase-medially.
Two analyses showed effects of the predictability of a neighboring word.
The word eigenlijk was more likely to be monosyllabic if it was preceded by a
more frequent word. This effect is in line with earlier findings by Pluymaekers,
Ernestus and Baayen (2005b) for this same word. In addition, we found an
effect of the frequency of occurrence of the following word on the likelihood of
creaky voice in the full vowel: creaky voice was more often present when the
word token was followed by words of higher frequencies.
These results with respect to creaky voice are interesting. We see that
creaky voice is more often absent under the same conditions where segments
tend to be absent and shorter (i.e. at higher speech rates and in phrase-medial
position). This suggests that the absence of creaky voice results from the same
mechanisms that also reduce segments. This hypothesis, however, does not fit
with our observation that creaky voice tends to be more often present if the
following word is of a higher frequency: creaky voice behaves differently in this
respect from segments, which tend to be more, rather than less, reduced before
high frequency words. We propose that these conflicting results are the
33
consequence of the ambiguous character of creaky voice. On the one hand,
creaky voice at the start of a vowel-initial word may function as a phonetic cue
to prosodic strength (cf. Jongenburger & van Heuven 1991), which tends to be
reduced in the same conditions where segments are reduced. On the other hand,
separate from this function in marking vowel-initial word onsets, the presence of
creaky voice may in some cases result from reduced articulatory effort in voicing
(cf. Gobl and Ní Chasaide 2003). In that case, we expect the presence of creaky
voice in those contexts where segments tend to be reduced, including before a
high frequency word. Further research is needed to test this hypothesis.
Unlike previous studies (e.g. Bell et al. 2003; Pluymaekers, Ernestus, and
Baayen 2005b; Bell et al. 2009), we did not find an effect of the frequency of the
following word on the presence versus absence of segments or on phone, affix,
or token durations. A possible explanation is that the previous studies did not
incorporate following word as a random variable. This hypothesis is supported by
an analysis showing that the likelihood of a monosyllabic token is predicted by
the frequency of the following word, if following word is not incorporated as a
random effect as well. The random effect of the following word was significant in
half of our analyses.
These observations may give some clues as to what may be stored in the
mental lexicon. On the one hand, the high frequency of highly reduced forms
and the existence of consistent phonetic patterns in their production encourages
the conclusion that multiple variants are stored as separate pronunciation
targets. On the other hand, such variants need to retain some link to the word’s
canonical form in order to account for the presence of articulatory residues in
the phonetic detail of the forms produced. Put differently, a reduced token of
eigenlijk, even if broadly transcribable as [ɛɪk], probably often differs from the
same phoneme string produced in the content word eik 'oak', as other authors
have demonstrated for word pairs like (reduced) support versus sport (e.g.
Manuel 1991; Manuel et al. 1992). To substantiate these suggestions requires
further acoustic analysis—in particular of spectral properties which we could not
address in this paper.
Furthermore, our data show that models of speech production have to
take the rhythm in the phrase into account when explaining the reduction degree
of eigenlijk. This strongly suggests that the assumption that degree of reduction
is only determined by how much time a speaker needs to plan and produce the
word or the following word is too simplistic: the speaker also appears to be
concerned with at least the rhythm of the phrase.
From a perceptual point of view, the data emphasize that the speech
comprehension system has to deal with a great deal of variation when it comes
to recognizing eigenlijk. One aspect that may help the listener is that several
core properties of eigenlijk are high in acoustic salience: the formant dynamics
characteristic of the diphthong /ɛɪ/, the compact mid-frequency spectral
prominences that characterize velar and uvular obstruents, and the strident
nature of uvular fricatives. These landmarks may guide the listener, enabling the
detection of finer details that cue the word’s identity, such as traces of /l/.
34
Perceptual experimentation is needed to test the role of the various gross and
subtle acoustic characteristics that we have identified.
7. Rethinking reduction
The research reported on in this chapter has generated data that may cast doubt
on some of the common assumptions about the phenomenon of speech
reduction. We showed that the variation in the pronunciation of discourse
markers may be substantial. The unreduced variant may be less common than
reduced variants (in our case only 20% of the tokens are unreduced). This raises
the question which form of such a word should be considered as canonical: the
full form, which is represented in orthography, or the most frequently occurring
reduced form.
Furthermore, our data show that, in contrast to what is generally
assumed, highly reduced discourse markers can occur in prosodically strong
positions. Reduction is therefore not for all words restricted to unaccented
positions. The occurrence of highly reduced forms in accented positions
underlines that reduced forms of at least some words are not special and can
occur without restrictions.
The pronunciation variation displayed by eigenlijk is conditioned, among
other factors, by the rhythm of the phrase, and shows large differences between
speakers. Moreover, a form may be reduced in one aspect, but not in another.
This strongly suggests that reduction is not a fully automatic process that arises
when speakers are under time pressure. Speakers clearly have a choice whether
to reduce and how to reduce, and they make this choice, among others, on the
basis of the rhythm of the phrase, while adhering to their own speech habits.
Finally, we found that every pronunciation variant of eigenlijk appears to
include two landmarks that may be considered to be the main characteristics of
the word (the full vowel and a velar/uvular consonant). These landmarks raise
questions about speech processing. Are these landmarks indicated in the mental
lexicon? How do listeners use these landmarks during word recognition?
We conclude that this corpus study has shown that many aspects of the
phenomenon of speech reduction are not yet well understood. We call for more
detailed qualitative and quantitative analyses of many tokens of individual words
produced in casual speech because these studies substantially extend our
knowledge about speech reduction and about speech production and perception.
Acknowledgements
We would like to thank Harald Baayen for facilitating the second author's stay at
the Max Planck Institute for Psycholinguistics, and Brechtje Post for her advice
on the prosodic analysis of our phrases. Furthermore, we would like to thank
Sarah Hawkins, Ellen Aalders and three anonymous reviewers for their helpful
comments on an earlier version of the paper. This research was partly funded by
a vici grant from the Netherlands Organization for Scientific Research.
35
References
Aalders, Ellen, Mirjam Ernestus. In preparation. Is there a link between schwa
reduction and the realization of following /p, t/ in casual speech?
Akaike, Htrotugu. 1974. A new look at the statistical model identification. IEEE
Transactions on Automatic Control 19, 716–723.
Aoyagi, Maki. In press. Japanese vowel devoicing and voicing control in the C/D
model. Journal of the Phonetic society of Japan.
Bell, Alan, Jason. M. Brenier, Michelle Gregory, Cynthia Girand & Dan Jurafsky.
2009. Predictability effects on durations of content and function words in
conversational English. Journal of Memory and Language 60. 92-111.
Bell, Alan, Daniel Jurafsky, Eric Fosler-Lussier, Cynthia Girand, Michelle Gregory
& Daniel Gidea. 2003. Effects of disfluencies, predictability and utterance
position on word form variation in English conversation. Journal of the
Acoustical Society of America 113(2). 1001-1024.
Bergen, Geertje van, Rik van Gijn, Lotte Hogeweg & Sander Lestrade. 2011.
Discourse marking and the subtle art of mind-reading: The case of Dutch
eigenlijk. Journal of Pragmatics 43. 3877-3892.
Booij, Geert. 1995. The phonology of Dutch. Oxford: Clarendon Press.
Cruttenden, Alan. 1994. Gimson’s Pronunciation of English (5th edition). London:
Edward Arnold.
Davidson, Lisa. 2006. Schwa elision in fast speech: segmental deletion or
gestural overlap? Phonetica 63. 79-112.
Ernestus, Mirjam. 2000. Voice assimilation and segment reduction in casual
Dutch, a corpus-based study of the phonology-phonetics interface. Utrecht:
Netherlands National Graduate School of Linguistics.
Ernestus, Mirjam & Natasha Warner. 2011. An introduction to reduced
pronunciation variants. Journal of Phonetics 39. 253-260.
Gahl, Susanne. 2008. "Thyme" and "Time" are not homophones. Word durations
in spontaneous speech. Language 84. 474-496.
Gahl, Susanne, Yao Yao & Keith Johnson. 2012. Why reduce? Phonological
neighborhood density and phonetic reduction in spontaneous speech. Journal
of Memory and Language 66. 789-806.
Gobl, Christer & Ní Chasaide, Ailbhe. 2003. The role of voice quality in
communicating emotion, mood and attitude. Speech Communication 40. 189-
212.
Gordon, Matthew, Paul Barthmaier & Kathy Sands. 2002. A cross-linguistic
acoustic study of fricatives. Journal of the International Phonetic Association
32. 141-174.
Guy, Gregory R. 1991. Explanation in variable phonology: An exponential model
of morphological constraints. Language Variation and Change 3. 1-32.
Hanique, Iris, Mirjam Ernestus & Lou Boves. 2015. Choice and pronunciation of
words: Individual differences within a homogeneous group of speakers.
Corpus Linguistics and Linguistic Theory 11. 161-185.
Johnson, Keith. 2004. Massive reduction in conversational American English. In
Kiyoko Yoneyama & Kikuo Maekawa (eds.), Spontaneous speech: Data and
36
analysis. Proceedings of the 1st session of the 10th international symposium,
29-54. Tokyo: The National International Institute for Japanese Language.
Jongenburger, Willy & Vincent van Heuven. 1991. The distribution of (word-
inital) glottal stop in Dutch. In Ans van Kemenade & Frank Drijkoningen
(eds.), Linguistics in the Netherlands, 101-110. Amsterdam: John Benjamins
Publishing Company.
Kelly, Michael H. & J. Kathryn Bock. 1988. Stress in time. Journal of
Experimental Psychology: Human Perception and Performance 14. 389-403.
Keune, Karen, Mirjam Ernestus, Roeland van Hout & R. Harald Baayen. 2005.
Social, geographical, and register variation in Dutch: From written mogelijk to
spoken mok. Corpus Linguistics and Linguistic Theory 1. 183-223.
Kohler, Klaus J. 1990. Segmental reduction in connected speech in German:
Phonological facts and phonetic explanations. In William J. Hardcastle & Alain
Marchal (eds.), Speech production and speech modelling, 21-33. Dordrecht:
Kluwer Academic Publishers.
Kohler, Klaus J. 1999. Articulatory prosodies in German reduced speech.
Proceedings of the XIVth International Congress of Phonetic Sciences 1. 89-
92.
Koreman, Jacques. 2006. Perceived speech rate: the effects of articulation rate
and speaking style in spontaneous speech. Journal of the Acoustical Society of
America 119. 582-596.
Labov, William L. 2001. Principles of linguistic change: Social factors. Oxford:
Blackwell Publishers.
Local, John. 2003. Variable domains and variable relevance: Interpreting
phonetic exponents. Journal of Phonetics 31(3). 321-339.
Manuel, Sharon Y. 1991. Recovery of deleted schwa. Perilus: Papers from the
Symposium on Current Phonetic Research Paradigms for Speech Motor
Control, University of Stockholm Institute of Linguistics. 115-118.
Manuel, Sharon Y. 1995. Speakers nasalize /∂/ after /n/, but listeners still hear
/∂/. Journal of Phonetics 23(4). 453-476.
Manuel, Sharon Y., Stefanie Shattuck-Hufnagel, Marie K. Huffman, Kenneth N.
Stevens, Rolf Carlson & Sheri Hunnicutt. 1992. Studies of vowel and
consonant reduction. Proceedings of the Second International Conference on
Spoken Language Processing. 943-946.
Niebuhr, Oliver & Klaus J. Kohler. 2011. Perception of phonetic detail in the
identification of highly reduced words. Journal of Phonetics 39(3). 319-329.
Oostdijk, Nelleke. 2000. The Spoken Dutch Corpus project. The ELRA newsletter
5(2). 4-8.
Phillips, Betty S. (1994). Southern English glide deletion revisited. American
Speech 69. 15-127.
Plug, Leendert. 2005. From words to actions: the phonetics of eigenlijk in two
communicative contexts. Phonetica 62. 131-145.
Pluymaekers, Mark, Mirjam Ernestus & R. Harald Baayen. 2005a. Lexical
frequency and acoustic reduction in spoken Dutch. Journal of the Acoustical
Society of America 118. 2561 - 2569.
37
Pluymaekers, Mark, Mirjam Ernestus & R. Harald Baayen. 2005b. Articulatory
planning is continuous and sensitive to informational redundancy. Phonetica
62. 146-159.
R Core Team. 2014. R: A Language and Environment for Statistical Computing.
Vienna: R Foundation for Statistical Computing.
Raymond, William D., Robin Dautricourt & Elizabeth Hume. 2006. Word-internal
/t,d/ deletion in spontaneous speech: Modeling the effects of extra-linguistic,
lexical, and phonological factors. Language Variation and Change 18. 55-97.
Son, Rob J. J. H. van & Louis C. W. Pols. 1990. Formant frequencies of Dutch
vowels in a text, read at normal and fast rate. Journal of the Acoustical
Society of America 88. 1683-1693.
Son, Rob J. J. H. van & Louis C. W. Pols. 1992. Formant movements of Dutch
vowels in a text, read at normal and fast rate. Journal of the Acoustical
Society of America 92. 121-127.
Strik, Helme, Joost van Doremalen & Catia Cucchiarini. 2008. Pronunciation
reduction: how it relates to speech style, gender, and age. Interspeech, 1477-
1480.
Torreira, Francisco & Mirjam Ernestus. 2011. Vowel elision in casual French: the
case of vowel /e/ in the word c'était. Journal of Phonetics 39, 50-58.
Warner, Natasha & Benjamin V. Tucker. 2011. Phonetic variability of stops and
flaps in spontaneous and careful speech. The Journal of the Acoustic Society
of America 130(3). 1606-1617.
38
Appendix
Variation dicussed in sections 4 and 5 for which we can provide the number of
tokens. The first two columns provide information about segmental variation, the
second two columns about subsegmental variation
Segmental variation Number
of tokens
(out of
159)
Subsegmental variation Number
of tokens
(out of
159)
Initial vowel is monoph-
tongal
6 Creaky voice for first vowel 83
First schwa is absent 139 Voicing of velar stop 25
Velar fricative is absent 22 At least partly devoiced /l/ 21
Second schwa is absent 58 At least partly devoiced
schwa
7
Velar stop is absent 19
